PrASD: Intelligent Multi-Agent Framework for Predictive-Based Abnormal State Detection in Smart Factory Devices

Source abstract

Effective predictive maintenance in smart factories is essential for enhancing reliability and optimizing production processes. This paper introduces PrASD, a multi-agent framework that integrates classification and regression techniques for advanced state detection and prediction. PrASD utilizes specialized deep learning agents to recognize operational states, model temporal dynamics, and generate anomaly scores. A key feature of PrASD is its ability to facilitate predicted-state classification by forecasting numerical trajectories while providing discrete interpretations (state classes) at specific time points. Feature engineering involves applying a two-dimensional wavelet transform to time windows of sensor values, facilitating the extraction of both low and high-frequency signals, which enhances the robustness of the classification process. In addition, the framework incorporates probabilistic anomaly detection that leverages prediction errors, as well as the disagreement and confidence levels of classification agents. This solution significantly enhances the effectiveness of anomaly detection. PrASD also coordinates hyperparameter optimization of the classifier, regressor, and anomaly detection modules. This approach mitigates performance degradation associated with independent tuning, ensuring cumulative improvements. By combining continuous predictions with discrete classifications, PrASD offers a comprehensive view of system health. The framework is adaptable for analyzing industrial system behavior across various known state categories, including undesirable states and scenarios focused solely on normal operations. Experimental evaluations on datasets from a condition monitoring device and a manufacturing robotic arm demonstrate that PrASD achieves high performance, underscoring its potential as a robust tool for predictive maintenance in complex environments.